Professional Documents
Culture Documents
Minimax Theorem: Appendix G
Minimax Theorem: Appendix G
Minimax Theorem: Appendix G
Minimax theorem
mmax.thm < 1> Theorem. Let K be a compact convex subset of a Hausdorff topological
vector space X, and C be a convex subset of a vector space Y. Let f be a
extend to real-valued function defined on K × C such that
R ∪ {∞} valued f ?
(i) x → f (x, y) is convex and lower-semicontinuous for each y ,
(ii) y → f (x, y) is concave for each x .
Then
infx∈K sup y∈C f (x, y) = sup y∈C infx∈K f (x, y).
Proof. Note that assumption (i) means precisely that K y,t := {x ∈ K :
f (x, y) ≤ t} is compact and convex, for each fixed t ∈ R and y ∈ C.
Clearly the left-hand side of the asserted equality is ≥ the right-hand
side. It therefore suffices to prove that if M is a real number for which M ≥
right-hand side then, for each > 0, the left-hand side is ≤ M + . The
assumption on M means that infx∈K f (x, y) ≤ M , and hence K y,M+ = ∅, for
every y. If we can prove that ∩ y∈Y K y,M+ = ∅ then there will exist an x0 for
which f (x0 , y) ≤ M + for every y, implying that sup y f (x0 , y) ≤ M + .
Replacing f by f − (M + ), we reduce to the case where M + = 0
and infx∈K f (x, y) ≤ − and K y,0 = ∅, for each y. We need to show that
∩ y∈Y K y,0 = ∅. Compactness of each K y,0 simplifies the task to showing that
∩ y∈Y0 K y,0 = ∅ for each finite subset Y0 of Y. An inductive argument will then
reduce even further to the case where Y0 := {y1 , y2 }, a two-point set, which is
the case that I consider first.
Abbreviate K yi ,0 to K i , and f (x, yi ) to f i (x). Thus K i = {x : f i (x) ≤ 0}.
For the purposes of obtaining a contradiction, suppose K 1 ∩ K 2 = ∅. The
contradiction will appear if we find a number α in [0, 1] for which
alpha.def <2> (1 − α) f 1 (x) + α f 2 (x) ≥ 0 for all x in K ,
for then the concavity of f (x, ·) would give the lower bound infx∈K f (x, yα ) ≥
0, where yα := (1 − α)y1 + αy2 .
Inequality <2> is trivial if x ∈ / K 1 ∪ K 2 , for then both f 1 (x) > 0 and
f 2 (x) > 0. For it to hold at each x in K 1 we would need
− f 1 (x1 )
lower.alpha <3> α ≥ sup .
x1 ∈K 1 2 1 ) − f 1 (x 1 )
f (x
Notice that the supremum on the right-hand side is ≥ 0. For inequality <2> to
hold at each x in K 2 we would need
f 1 (x2 )
upper.alpha <4> α ≤ inf .
x2 ∈K 2 f 1 (x 2 ) − f 2 (x 2 )
f1
f2
x1 xθ x2
The inequality <5> is trivial when either f 1 (x1 ) = 0 or f 2 (x2 ) = 0.
We need only consider a pair with f 1 (x1 ) < 0 and f 2 (x2 ) < 0. Define θ
in (0, 1) as the value for which (1 − θ) f 1 (x1 ) + θ f 1 (x2 ) = 0, then define
xθ := (1 − θ )x1 + θ x2 . By convexity, f 1 (xθ ) ≤ 0, implying that xθ ∈ K 1 and
(1 − θ ) f 2 (x1 ) + θ f 2 (x2 ) ≥ f 2 (xθ ) > 0. Thus
− f 1 (x1 ) θ f 2 (x1 )
= < ,
f 1 (x2 ) 1−θ − f 2 (x2 )
which gives <5>.
Existence of an α satisfying constraints <3> and <4> now follows, which,
via the contradiction, lets us conclude that K 1 ∩ K 2 = ∅.
To extend the conclusion to an intersection of finitely many sets K yi ,0 , for
i = 1, 2, . . . , m, first invoke the result for pairs to see that K i := K y1 ,0 ∩K yi ,0 = ∅
for i = 2, . . . , m, then repeat the pairwise argument with f restricted to
K y1 ,0 × C. And so on. After m − 1 repetitions we reach the desired conclusion,
that ∩i≤m K yi ,0 = ∅.
convex.hull <6> Example. Let P and Q be collections of probability measures on the same
space (X, A). Write co(P) and co(Q) fortheir convex hulls. That is, co(P)
consists of all finite linear combinations i αi Pi , with αi ≥ 0 and i αi = 1
and Pi ∈ P, with a similar definition for co(Q). Write T for the collection of
all tests, that is, functions ψ in M(X, A) for which 0 ≤ ψ ≤ 1. If ψ ∈ T, the
function ψ̄ := 1 − ψ is also a test.
If both P and Q are dominated by some sigma-finite measure λ, the
Minimax Theorem will show that
lecam.mmax <7> inf sup{Pψ + Qψ̄ : P ∈ P, Q ∈ Q} = sup{
:
P ∧ Q
∈ co(Q)}.
P ∈ co(P), Q
ψ∈T
Before proving the equality, first note that the left-hand side does not change if
we replace both P and Q by their convex hulls, because
i
αi Pi ψ + j
β j Q j ψ̄ = i, j
αi β j Pi ψ + Q j ψ̄ .
Thus there is no loss of generality in assuming both P and Q are convex. The
set C of all signed measures of the form ν = P − Q, with P ∈ P and Q ∈ Q, is
[§] 2. Notes
Sources: Kneser (1952)? Fan (1953)? Sion (1958)? See Millar (1983, page 92).
Possibly also some handwritten notes from other lectures by Millar.
Le Cam (1973) stated the result from Example <6> with a reference to
Kraft (1955) for the proof. Kraft attributed the result and the proof to Le Cam.
When the families are not dominated, the proof breaks down. Le Cam (1986,
page 476) stated generalized forms of the result, based on compactifications of
either T or C.
References